CODEWITHSUNDEEP
numpypytorchtorchunfold
Khawar Islam
Khawar Islam

Reputation: 2914

"im2col_out_cpu" not implemented for 'Byte'

I am trying to generate overlap patches from image size (112,112) but i am unable to do so. I have already tried a lot but it didn't work out.

**Code**
import torch
import numpy as np
import torch.nn as nn
from torch import nn
from PIL import Image
import cv2
import os
import math
import torch.nn.functional as F
import torchvision.transforms as T
from timm import create_model
from typing import List

import matplotlib.pyplot as plt
from torchvision import io, transforms
from utils_torch import Image, ImageDraw
from torchvision.transforms.functional import to_pil_image

IMG_SIZE = 112
# PATCH_SIZE = 64

resize = transforms.Resize((IMG_SIZE, IMG_SIZE))
img = resize(io.read_image("Adam_Brody_233.png"))
img = img.to(torch.float32)

image_size = 112
patch_size = 28
ac_patch_size = 12
pad = 4

img = img.unsqueeze(0)
soft_split = nn.Unfold(kernel_size=(ac_patch_size, ac_patch_size), stride=(patch_size, patch_size), padding=(pad, pad))
patches = soft_split(img).transpose(1, 2)

fig, ax = plt.subplots(16, 16)
for i in range(16):
    for j in range(16):
        sub_img = patches[:, i, j]
        ax[i][j].imshow(to_pil_image(sub_img))
        ax[i][j].axis('off')

plt.show()

Traceback

Traceback (most recent call last):
  File "/home/cvpr/Documents/OPVT/unfold_ours.py", line 32, in <module>
    patches = soft_split(img).transpose(1, 2)
  File "/home/cvpr/anaconda3/envs/OPVT/lib/python3.7/site-packages/torch/nn/modules/module.py", line 727, in _call_impl
    result = self.forward(*input, **kwargs)
  File "/home/cvpr/anaconda3/envs/OPVT/lib/python3.7/site-packages/torch/nn/modules/fold.py", line 295, in forward
    self.padding, self.stride)
  File "/home/cvpr/anaconda3/envs/OPVT/lib/python3.7/site-packages/torch/nn/functional.py", line 3831, in unfold
    _pair(dilation), _pair(padding), _pair(stride))
RuntimeError: "im2col_out_cpu" not implemented for 'Byte'

Upvotes: 0

Views: 1063

Answers (1)

darth baba
darth baba

Reputation: 1398

Yes this is an open issue in PyTorch. A simple fix is just to convert your image tensor from ints to floats you can do it like this:
img = img.to(torch.float32)
This should solve your problem

Upvotes: 0

Related Questions